JSHookMCP: 735 reverse-engineering tools for AI agents in 3K tokens

Every reverse-engineering tool I tried lived on exactly one layer. A browser tool can read the DOM but not the V8 JIT that produced it. A debugger sees the process but not the HTTP/2 frame that triggered it. The analysis stayed manual, and an agent could not follow one signal across those boundaries. jshookmcp started as a browser/CDP toolkit and grew downward: network and TLS keylog, DOM, JS AST, WASM, Mojo IPC, native memory, then an in-process ARM64 interpreter for Android JNI, then syscalls. The point is that a finding on one layer can be linked back to the layer that caused it, in one evidence graph instead of five terminals. The hard part turned out to be context. 735 tools is roughly 109K tokens of schema, which no agent can afford. So the default profile exposes 8 meta-tools and a local search tier (BM25 + ONNX embeddings + RRF), and domains activate on demand with a 30-minute TTL. That lands at about 1.9K tokens. It is AGPL-3.0. Ask me anything about the JNI interpreter or how the cross-domain graph is assembled.